CN101889343B - Hybrid microscale-nanoscale neuromorphic integrated circuit - Google Patents
Hybrid microscale-nanoscale neuromorphic integrated circuit Download PDFInfo
- Publication number
- CN101889343B CN101889343B CN2008801196638A CN200880119663A CN101889343B CN 101889343 B CN101889343 B CN 101889343B CN 2008801196638 A CN2008801196638 A CN 2008801196638A CN 200880119663 A CN200880119663 A CN 200880119663A CN 101889343 B CN101889343 B CN 101889343B
- Authority
- CN
- China
- Prior art keywords
- signal
- integrated circuit
- neuron
- input
- computing unit
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
- G06N3/065—Analogue means
Abstract
Embodiments of the present invention include hybrid microscale-nanoscale neuromorphic integrated circuits that include an array of analog computational cells fabricated on an integrated-circuit-substrate. The analog electronic circuitry within each computational cell connected to one or more pins of a first type and to one or more pins of a second type that extend approximately vertically from the computational cells. The computational cells are additionally interconnected by one or more nanowire-interconnect layers, each nanowire-interconnect layer including two nanowire sublayers on either side of a memristive sublayer, with each nanowire in each nanowire sublayer of an interconnect layer connected to a single computational-cell pin and to a number of nanowires in the other nanowire sublayer of the interconnect layer.
Description
Draw on the side of association request
The application requires the provisional application No.60/992 of submission on December 5th, 2007,663 rights and interests.
Technical field
The application relates to material science and circuit application, relates in particular to the hybrid microscale level-nanoscale integrated circuit architecture that is used to realize various complicated circuit, treatment system and calculation element, comprises the neuromorphic circuit of simulating biological neuron circuit.
Background of invention
The current method that is used to realize the micron order circuit of computer processor, memory and other computing machine has caused current densities and the noticeable exponential growth of computing capability since half a century in the past.Yet; Development is summarized and the disposal ability and the characteristic density double surge every two years that are called as " Moore's Law " begun to slow down according to computer for many years; Nowadays run into physical restriction and practice constraint because further reduce characteristic size, comprise along with signal line dimensions reduce and make resistance increase, since reduce along with characteristic size the electric capacity of characteristic increase produces heat increase the difficulty that the heat that makes from processor dispels the heat increasing, owing to the more and more littler difficulty that characteristic ran into of manufacturing causes the higher defective of processor and memory device and failure rate and in order further to reduce the difficulty that characteristic size manufactures and designs facility and method.The confirmation that further reduces of characteristic size becomes more and more difficult in the integrated circuit, has begun to adopt the multiple alternative method of increase based on the computing capability of the electronic device of integrated circuit.As an example, processor supplier is making polycaryon processor, and it has increased computing capability through Distribution Calculation on a plurality of kernels of executed in parallel multiple-task.Other effort comprises uses various molecular electronic techniques on nanoscale, to make circuit, and through to improve fault transmission fix the defect and the reliability problem of data-signal through electronic communication media with the theoretical method that uses the similar mode of error correcting code to use based on information science.Other effort is to the nanometer circuit that also is called as " neuromorphic circuit " of the biological neuron circuit of research and development simulations, biological neuron circuit to biological organic organization provide that efficient is considerable, low-power, parallel calculation mechanism.Yet; Many current methods adopt the conventional logic that realizes in complementary metal oxide semiconductors (CMOS) (CMOS) technology to realize cynapse neuromorphic circuit equivalent thing; This has seriously limited the density that neuron neuromorphic circuit equivalent thing can be made, and is limited in generally that every square centimeter of semiconductor chip surface is long-pending to have several thousand neurons.Therefore the researcher of neuromorphic circuit and research staff find, in order to produce enough intensive neuromorphic circuit, need new technology and new framework.
Summary of the invention
Embodiments of the invention comprise hybrid microscale level-nanoscale neuromorphic integrated circuit, and these integrated circuits are included in the analogue computer array of making on the integrated circuit substrate.Analog circuit in each computing unit is connected to one or more first kind pins and one or more second type of pin, and they basically vertically stretch out from computing unit.Computing unit is in addition through one or more nanowire interconnections layer interconnection; Each nanowire interconnections layer comprises two nano wire sublayers on the either side of memory resistor property (memristive) sublayer, wherein every nano wire in each nanometer sublayer of interconnection layer is connected to single computing unit pin and is connected to several nano wires in another nano wire sublayer of interconnection layer.Micron order of the present invention-nanoscale composite nerve form circuit framework can be used for realizing very a large amount of different complex electronic circuit, computing system and calculation element, comprises neuromorphic stratiform cortex circuit.
The accompanying drawing summary
Fig. 1 illustrates the basic calculating unit of the hybrid microscale level-nanoscale neuromorphic integrated circuit of expression one embodiment of the invention.
Fig. 2 is illustrated in the various embodiments of the present invention memory resistor between two nano wires of cynapse behavior modeling is tied.
Fig. 3 A-3B is illustrated in the essence electronic characteristic that is used in the various embodiments of the present invention to the memory resistor of cynapse modeling knot.
Fig. 4 is illustrated in the nerve cell that serves as the basic calculating unit among each embodiment of hybrid microscale level-nanoscale neuromorphic integrated circuit of the present invention.
Fig. 5 A-E illustrates the inner working of a simple illustration nerve cell of the expression embodiment of the invention.
Fig. 6 illustrates the general internal circuit according to the nerve cell of various embodiments of the present invention.
Fig. 7 is illustrated in the custom transmission gate computing unit that adopts in hybrid microscale level-nanoscale neuromorphic integrated circuit of representing the embodiment of the invention.
Fig. 8 illustrates the remodeling of the custom transmission gate computing unit of input signal according to an embodiment of the invention.
Fig. 9 illustrates the example of the computing unit of two kinds of addition type that adopted in hybrid microscale level-nanoscale neuromorphic integrated circuit of representing the embodiment of the invention, comprise input computing unit and output computing unit.
Figure 10 illustrates the input and output signal of the input computing unit of micron order-nanoscale composite nerve form circuit of representing one embodiment of the invention.
Figure 11 A-B is illustrated in the interconnection of the computing unit in hybrid microscale level-nanoscale neuromorphic integrated circuit of representing one embodiment of the invention.
Figure 12 A-F illustrates the manufacturing of hybrid microscale level-nanoscale neuromorphic integrated circuit of representing one embodiment of the invention.
Figure 13 illustrates the two-way signaling through the nano wire transmission of the nanowire interconnections layer of hybrid microscale level-nanoscale neuromorphic integrated circuit of representing one embodiment of the invention.
Figure 14 illustrates the electricity of the memory resistor nanowire-junction of neural and postsynaptic neuron of the presynaptic that is used to interconnect and leads variation.
Figure 15 illustrates the realization of six computing unit dipoles according to an embodiment of the invention, and it shows the simulation behavior similar with digital logic flip-flops.
Figure 16 illustrates many computing units module of second type that can in representing the hybrid microscale level-nanoscale neuromorphic integrated circuit of one embodiment of the invention, adopt.
Figure 17 illustrates the layering interconnection of the computing unit in hybrid microscale level-nanoscale neuromorphic integrated circuit of representing one embodiment of the invention.
Figure 18 illustrates the logical construction of second type that can in representing the hybrid microscale level-nanoscale neuromorphic integrated circuit of one embodiment of the invention, realize.
Figure 19 illustrates can be constructed in accordance and can adopt to realize the stratiform cortex circuit module of stratiform cortical layer neuromorphic circuit with a large amount of adjacent stratiform cortical layer modules.
Detailed Description Of The Invention
Embodiments of the invention are to providing extremely intensive, low-power consumption analog neuron form circuit for realizing the hybrid microscale level-nanoscale neuromorphic integrated circuit of different electric electronic circuit, computing system and the calculating device of unlimited amount in fact.Term " micron order " refers to the micron order and submicron order assembly and the characteristic that---comprise photoetching process, doping, etching and linearisation---and in based on the integrated circuit layer of silicon or the integrated circuit layer based on other semiconductor substance, realize through the multiple technologies of making integrated circuit, and it has the minimum dimension down to 10 and 100 nanometers from several microns.Term " nanoscale " refer to use make nano wire and comprise minimum dimension that the whole bag of tricks of the various nano thread structures of netted and chi structure realizes from tens nanometers down to littler characteristic and assembly less than 10 nanometers.
Biological neuron circuit and the neuromorphic circuit that manufacturing is used for simulating biological circuit based on through cynapse respectively with the nerve cell and the neural computing unit of other nerve cell cell interconnection.Neuronic activity level generally is applicable to short-term storage, generally be applicable to the memory in mid-term as prolonging to stimulate with other inhibition of being accustomed to the nervous excitation of response results, and synapse weight is applicable to long term memory.In general, in biological neuron circuit, given neuron can be through up to 10000 or more a plurality of cynapse and other neuron interconnections.That kind as mentioned above; Because the considerable part expense of integrated circuit is used to simulate the cynapse function in the submicron order logical circuit, therefore the trial of moulding the neuromorphic circuit of micron order from integrated circuit and submicron order electronic building brick generally causes quite low neuron density up to now.Owing to need so more cynapse than neuron, in many neuromorphic integrated circuit designs, the most surfaces of neuromorphic integrated circuit is devoted between the neuron of quantity much less, to realize that a large amount of cynapses connect.
Embodiments of the invention are tied the neuromorphic integrated circuit that high neuron density is provided through the memory resistor that cynapse is embodied as between the nano wire.Cynapse and simulate dendron and aixs cylinder in the biological neuron circuit through the nanowire signal line of cynapse interconnection; And make in the nano interconnection layer on the semiconductor integrated circuit layer, keep surface of semiconductor integrated circuit is called " nerve cell " for being implemented in hereinafter neuron computes unit and many computing units module thus.Therefore; Hybrid microscale level-nanoscale neuromorphic integrated circuit of design and manufacturing adopts memory resistor nanowire-junction rather than logical circuit to realize cynapse according to the present invention; And realize in the nanowire interconnections layer that is interconnected in semiconductor integrated circuit layer top based on cynapse between cynapse and the nerve cell, thereby very large neuron density is provided with three-dimensional hybrid micron order-nanoscale neuromorphic circuit framework.
Fig. 1 illustrates the basic calculating unit of hybrid microscale level-nanoscale neuromorphic integrated circuit of representing one embodiment of the invention.This computing unit comprises the formula area of the semiconductor integrated circuit layer 102 that four conductive pin 104-107 therefrom vertically stretch out.Interconnect to conductive pin such as the horizontal nanowire of nano wire among Fig. 1 108 pad class formation through for example pad class formation 110, and cross a plurality of computing units along the two-dimensional array of the computing unit of hybrid microscale level-nanoscale nano shape integrated circuit at the neighborhood interior span of computing unit 102 and extend point-blank.As hereinafter further describe, the semiconductor integrated circuit layer of computing unit 102 comprises various interconnection and the simulated assembly of realizing neuron models or other basic calculating device, wherein certain some with explanation in more detail hereinafter.Four vertical pins 104-107 are used for the simulated assembly in the semiconductor integrated circuit layer segment of computing unit 102 and circuit interconnection to the nano wire layer such as nano wire 108.Nano wire then can interconnect to contiguous computing unit with computing unit through nano wire and the memory resistor knot to the cynapse modeling.
Fig. 2 is illustrated in the various embodiments of the present invention memory resistor between two nano wires of cynapse behavior modeling is tied.In Fig. 2, first computing unit 202 is illustrated as to be positioned at and adjoins neighborhood calculation unit 204 parts.First nano wire 206 is connected to the vertical pins 208 of adjoining neighborhood calculation unit 204.Second nano wire 210 is electrically connected to the vertical pins 212 of computing unit 202, shown in the prospect of Fig. 2.First nano wire 206 and second nano wire 210 overlap each other in the zone that the little broken circle 214 Fig. 2 limits, and this crossover region is exaggerated in illustration 216.Between first nano wire 206 and second nano wire 210, have a substratum memory resistor material 218, it makes first nano wire and the second nano wire electric interconnection.Article two, the memory resistor between nano wire knot can be shown in illustration 218 such memory resistor mark 220 that symbolically is expressed as interconnection two signal line 222 and 224.As hereinafter further specifying ground, every nano wire in the interconnection layer can be through memory resistor knot and many different nanowire interconnections.
Fig. 3 A-B is illustrated in the essential characteristic electron that is used in the various embodiments of the present invention to the memory resistor of cynapse modeling knot.Fig. 3 A and 3B all illustrate the current/voltage curve chart of memory resistor knot.Trunnion axis 302 is represented voltages and vertical axis 304 expression electric currents.Voltage scanning shown in Fig. 3 A.The continuous voltage that comprises voltage scanning changes by voltage path 312 expressions, this voltage path 312 relatively with Fig. 3 A in current/voltage curve chart 316 align and thereunder second voltage axis, 314 draftings.Shown in Fig. 3 A; Through by realizing voltage scanning from the steady growth voltage of no-voltage 306 to voltage
, afterwards voltage be decreased to continuously negative voltage
afterwards voltage go back up to 0 (306 among Fig. 3 A).How the current/voltage curve changes during voltage scanning if illustrating the memory resistor conductivity of electrolyte materials.
At first; The memory resistor material is in low electricity and leads state; So that the current amplitude maintenance is relatively low; In the first of curve chart 318, (Fig. 3 A 306) increases to just under
from 0 along with voltage.Near
; Because the resistance of memory resistor material sharply descends or conductivity increases with non-linear form, electric current begins to rise rapidly 320.Remain height owing to drop to
memory resistor conductivity of electrolyte materials from
after the voltage, as the viewed that kind of being transmitted through the memory resistor material to the relevant voltage value in curve chart 322 and 324 parts of big relatively amplitude electric current.Near negative voltage
, the electricity of memory resistor material is led and is begun sharply to reduce 326 suddenly.The memory resistor material is in low electricity and leads state at
, and when zero increases, keeps low electricity to lead (328 among Fig. 3 A) once more at voltage.Shown in Fig. 3 B, second voltage scanning 330 is led with respect to the electricity that during first voltage scanning, produces the electricity of memory resistor material and is led increase, and this is by dotted line 332 expressions.Auxiliary voltage scanning can further make the electricity of memory resistor material lead with respect to the electricity that during last voltage scanning, produces and lead increase.Therefore, the memory resistor material shows non-linear that electricity leads under the voltage that is applied that increases continuously or reduce, and further shows the storage that previous electricity is led state.In other words, for employed all kinds memory resistor material in the embodiment of the invention, the physical state of memory resistor material w changed as the function of the current physical state of memory resistor material and the voltage that applies with respect to the time:
Current i through the memory resistor in embodiment of the invention knot be the function of the voltage that applies and conductivity of material,
The wherein electric g that leads is the current state of memory resistor material and the function of the voltage that applies:
i＝g(w，v)V
Shown in Fig. 3 A-B, the electricity of memory resistor knot is led the history that depends on the voltage that applies in the current voltage that applies and the last time interval.
Cynapse generally causes amplification or the decay that causes the signal of postsynaptic neuron j by presynaptic neuron i generation and through cynapse.In other model, the gain of cynapse or weight be in the scope of 0.0-1.0, the complete attenuation of 0.0 representation signal that wherein gains and 1.0 representatives that gain do not have signal attenuation.In these models, neuron has activity, and works as the active x of neuron i
iDuring greater than a threshold value, neuron sends the output signal.The Mathematical Modeling of neuron behavior is provided in the paragraph of back.Gain Z with the cynapse of presynaptic neuron i and postsynaptic neuron j interconnection
IjA Mathematical Modeling of rate of change be represented as:
Z wherein
IjBe weight or the gain that is produced through cynapse ij with presynaptic neuron i and postsynaptic neuron j interconnection;
ε is a pace of learning;
ω forgets speed;
x
iIt is the activity of neuron i;
x
jIt is the activity of neuron j;
F ((x
j) (ω z
Ij+ g (x
i))) be that variable is (x
j) (ω z
Ij+ g (x
i)) nonlinear function;
G (x
i) be the nonlinear function of the activity of neuron i; And
T is the time.
In many embodiment of the present invention, f () and g () are basic ∑ shapes.Exemplary ∑ shape or serpentine function are tanh ().When presynaptic neuron and postsynaptic neuron all have high activity, gain Z
IjIncrease rapidly.As item-ω z
IjSize greater than postsynaptic neuron g (x
i) the currency of nonlinear function of activity the time, item-ω z
IjGuarantee that the cynapse gain reduces in time.Since with the synapse weight of cynapse near 1.0 be used for reducing synapse weight and with the synapse weight of cynapse near 0.0 produce more and more littler feedback feedback term-ω z
Ij, the weight of cynapse can't increase or reduces with unrestricted mode.The Mathematical Modeling of cynapse behavior depends on the Mathematical Modeling of neuronal activity, and these models provide the mutual feedback to each other.As can through the Mathematical Modeling of cynapse gain is compared with the expression formula of the conductivity variations of top description memory resistor knot viewed; Particularly; Electricity derived function g (w; The electricity that v) is the memory resistor knot is led the physical embodiments that gain function can be provided, and its time derivative is expressed as above-mentioned Mathematical Modeling, because the neuronal activity f (x of cynapse model
i) and g (x
i) nonlinear function be relevant to the physics voltage between the neuron, and the gain z
IjPreset time spot correlation in the history of the voltage that memory resistor knot is applied.The function table dyne that the electricity of memory resistor property nanowire-junction is led this depend on the presynaptic that connects by the memory resistor nanowire-junction and postsynaptic neuron current activity and memory resistor nanowire-junction apply voltage history recently.Therefore, in each embodiment of the present invention, the memory resistor nanowire-junction of interconnected nanowires provides to make and as above-mentioned Mathematical Modeling is expressed, is suitable for the physical features that the current signal to the cynapse behavior modeling passes through.
Fig. 4 is illustrated in the nerve cell that serves as the basic calculating unit among each embodiment of hybrid microscale level-nanoscale neuromorphic integrated circuit of the present invention.Nerve cell is a kind of computing unit in hybrid microscale level-nanoscale neuromorphic integrated circuit.As stated, nerve cell 402 comprises four vertical conduction pin 404-407.These pins guide through its compass heading, and wherein compass graph 410 illustrates on the right side of Fig. 4 computing unit.NW pin 404 will conduct to from the output signal of nerve cell and the nano wire of NW pin 404 with 405 interconnection of SE pin with SE pin 405.SW pin 406 conducts to nerve cell 402 with the signal that NE pin 407 all will input to pin from the nano wire that is connected to these pins.SW pin 406 is conducted into nerve cell with the inhibition signal, and NE pin 407 conducts to nerve cell with the excitant input signal.The excitant input signal tends to increase neuronic activity, and the inhibition signal tends to reduce neuronic activity.
Basic nerve cell shown in Figure 4 402 is general realizes a kind of in the neuronic multiple different mathematics.Generally speaking, when the frequency and the number of the excitant signal that receives significantly surpassed frequency and the number of inhibition signal, neuronic activity generally increased on the threshold value activity value, and neuron sends the output signal through output pin 404,405 at this moment.
Input stimulus property signal is to receive through cynapse class memory resistor nanowire-junction other nerve cell from hybrid microscale level-nanoscale neuromorphic integrated circuit with input inhibition signal, and guides other computing unit of hybrid microscale level-nanoscale neuromorphic integrated circuit into through cynapse class memory resistor nanowire-junction by the output signal that nerve cell 402 sends.Nerve cell and neuromorphic circuit generally comprise various feedback mechanism, and show the non-linear behavior of each neuronic activity in control and the constraint neuromorphic circuit.Also can show the suitable sophisticated functions that can't use the modeling of closed type mathematic(al) representation usually and be difficult in traditional Digital Logical Circuits, realize even only comprise the median size neuromorphic circuit of the relatively small amount nerve cell through the intensive interconnection of cynapse based on Boolean logic.In Fig. 4; Input 412 and output 412 expressions: except receiving signal and send the signal through four vertical pins, nerve cell all can be through additional micron order or submicron order holding wire and the neighborhood calculation cell interconnection of in the semiconductor integrated circuit aspect of hybrid microscale level-nanoscale neuromorphic integrated circuit, realizing.
Fig. 5 A-5E illustrates the built-in function of a simple illustration nerve cell representing the embodiment of the invention.Shown in Fig. 5 A, whole excitant inputs (such as excitant input 502) of NE input pin (407 among Fig. 4) are sued for peace through sum operation 504.Equally, sued for peace through similar summing function 508 such as all inhibition input signals of inhibition input signal 506.From the additional input of semiconductor integrated circuit layer, such as the input among Fig. 4 412, can be included in the sum operation, maybe can be input in the functional unit of carrying out later stage nerve cell processing.Summation component 504,508 can be the input pin that is connected with the input nanowire signal line simply, maybe can comprise the amplifier module as other electronic building brick.Subsequently excitant signal sum and inhibition signal sum are input to signal generation functional unit 510, this functional unit produces the analog voltage signal by 512 expressions of the signal/time graph among Fig. 5 A.
Shown in Fig. 5 B, exemplary nerve cell comprises the leaky functional unit, and it is before the current time and comprise in certain time interval of current time the signal (shown in Fig. 5 A) that is produced by signal systematic function or operation is carried out continuous integral.Shown in Fig. 5 B, leak integration and can be regarded as time window function 520 is superimposed upon on the analog signal 522 and to being lower than these time window function 520 that part of analog signals and carry out integration.In Fig. 5 B and subsequent drawings, for ease of explanation, analog signal be expressed as without exception on the occasion of, yet, in fact analog signal can be on the occasion of, 0 or negative value.
Shown in Fig. 5 C, leak the integrating function assembly with integration after signal 528 export to and get threshold function assembly 530.When signal behind the integration have greater than as when getting the value of the definite threshold value of threshold function assembly institute, nerve cell sends one and exports signal, it is expressed as the spike wave train 532 shown in Fig. 5 C.Perhaps, nerve cell does not send signal, as the curve of the permanent null function 534 among Fig. 5 C is represented.Get the threshold function assembly and can be considered to the sense activation signal generating assembly, or alternatively comprise output signal generating assembly.
The operation of nerve cell can be regarded as convolution algorithm 540 to analog signal 542 and time window function Ψ 544 to produce through convolution function 546 shown in Fig. 5 D.Threshold value 548 is superimposed upon then on convolution function 546, or will be through downward translation one threshold distance of convolution function, to produce corresponding activity functions 550.Therefore, in the leakage integration of input signal produced the time interval (such as the time interval 552 and 554) of the value that is higher than threshold value, nerve cell was activated, shown in the activity curve 550 of nerve cell among Fig. 5 D.
Fig. 5 E summarizes the built-in function according to the nerve cell of various embodiments of the present invention.Excitant input 560 is by summation 562 and be input to signal generation function 564.Equally, inhibition signal 566 is sued for peace through summing function 568 and is inputed to signal generation function 564.Operation takes place signal can generate analog signal based on the input of making up, suing for peace according in the multiple Mathematical Modeling of neuron behavior any.For example, the signal generating function can be applied to excitant input signal and the inhibition input signal through suing for peace through summation with nonlinear function, and the result of linear combination nonlinear function application subsequently.Signal generating function 564 exports analog signal to leaky functional unit 570; This leaky functional unit 570 in the last time interval with integration after signal export to and get threshold function assembly 572, the said threshold function assembly 572 of getting judges whether that nerve cell has enough activity and sends the output signal.
As mentioned below, leaking integration is the special case of neuron circuit, and in the example of Fig. 5 A-5E, is shown exemplary neuron circuit because of its relatively easy illustration.In addition, can adopt with better function and neuron circuit flexibly, comprise and realizing as Huo Qijin-Huxley (Hodgkin-Huxley) neuron models that hereinafter is described and the circuit of shunting model.What should stress is, more powerful and flexibly in the neuron circuit in these functions, can not need get threshold component.Threshold device is useful when producing as the signal between the neuron when peaked wave in some neuromorphic circuit is listed as; Yet the generation of signal peaks need not by getting threshold component in other model, or alternatively signal can be more continuous and do not comprise the spike wave train.
Fig. 6 illustrates the general internal circuit of the nerve cell of a plurality of embodiment according to the present invention.Nerve cell 602 via output pin 604,606 through by the neighborhood calculation unit that signal is exported to the neuromorphic circuit such as the represented memory resistor property cynapse class nanowire-junction of the memory resistor mark of memory resistor mark 608.Nerve cell receives the inhibition signal and receives the excitant signal through second input pin 612 through first input pin 610.Pin assignment to output signal, inhibition input signal and excitant input signal among Fig. 6 is similar to the situation that Fig. 4 describes; Although can make different configurations in an alternative embodiment, and given hybrid microscale level-nanoscale neuromorphic integrated circuit can adopt multiple different pin configuration to computing unit.In nerve cell shown in Figure 6; Simple summing circuit 614,616 based on operational amplifier is used for excitant input signal and inhibition input signal are sued for peace to produce excitant signal input signal 620 and the inhibition signal input signal 622 through suing for peace through summation; They are imported into the internal circuit of nerve cell 624, and this internal circuit is realized combining signal generation, the integration of Fig. 5 A-5E description in the preceding text and getting threshold function.Accurate characteristic to leaking integration, signal is synthetic and gets the circuit that threshold value adopts depends on the characteristic of nerve cell institute receiving inputted signal and the accurate mathematical model that nerve cell is realized.For example, can be used for leaking integration, and that polytype summing circuit based on operational amplifier can be used to two of linear combinations is synthetic to carry out signal through summation input signal 620,622 based on the integrated circuit of operational amplifier and capacitor.
The nerve cell Mathematical Modeling that has number of different types.In given arbitrarily neuromorphic circuit, can adopt the some dissimilar nerve cell of realizing different neuron mathematics models correctly to simulate or to realize required Premium Features.A kind of Mathematical Modeling of nerve cell can be expressed as:
Wherein i is the neuron through modeling;
x
iIt is the activity of neuron i;
T is the time;
f
j(x
j) be the active nonlinear function of neuron j;
z
JiIt is the weight of the cynapse between neuron j and the i;
f
j(x
j) be the nonlinear function of the activity of neuron j;
N is the neuron number that links through excitant input and node i;
M is the neuron number that links through inhibition input and node i;
A and B are constants.
In this model, the activity of specific nerve cell i is first passive attenuation term-Ax
iLinear combination, second reflection excitant input through the feedback term correction with
And the 3rd reflection by the inhibition input of feedback term correction with
Under this model, the active x of neuron i
iScope be 0.0-1.0.When the neuronal activity height promptly near 1.0 the time, the feedback term in the equation is used for the further increase and the pressure of restricted activity makes active decline.On the other hand, as the active x of neuron i
iLow promptly near 0.0 o'clock, feedback term significantly reduces and nerve cell responds with muting sensitivity the inhibition input with the high sensitivity response the excitant input, increases the activity of nerve cell thus.As stated, when the activity of nerve cell is higher than threshold value, nerve cell will be exported signal through cynapse class memory resistor nanowire-junction and be transmitted into the neighborhood calculation unit, and also can export signal to the neighborhood calculation unit through semiconductor integrated circuit layer signal line.As stated; Many different possible Mathematical Modeling to the nerve cell operation is feasible; And can in single neuromorphic circuit, adopt many different model, wherein the nerve cell of the interior Different Logic aspect of layering neuromorphic circuit is realized the different mathematics of nerve cell behavior.
Fig. 7 is illustrated in the custom transmission gate computing unit that adopts in hybrid microscale level-nanoscale neuromorphic integrated circuit of representing the embodiment of the invention.As shown in Figure 7, custom transmission gate unit (HTG) does not adopt the vertical pins that nerve cell is connected to other nerve cell through cynapse class memory resistor nanowire-junction.On the contrary, HTG 702 receives input 704, also therefrom generation output 706 through the neighborhood calculation unit of semiconductor integrated circuit holding wire from the neuromorphic circuit.Compare the influence of nerve cell in response to inhibition and excitant input generation, HTG produces the more influence of long duration relatively to input signal.Fig. 8 illustrates and is accustomed to the correction of transmission gate computing unit to input signal according to an embodiment of the invention.In Fig. 8, input signal is drawn into signal strength signal intensity with respect to first curve 802 of time, and corresponding output signal is drawn into second curve 804 of signal strength signal intensity with respect to the time.Two curves among Fig. 8 respectively with respect to the time axle 806,808 alignment.When input signal 810 comprised positive electricity voltage crest 812, output signal 814 also comprised the positive electricity voltage crest 812 corresponding positive electricity voltage crests 816 with input signal.Yet in the output signal, voltage generally forms in crest and a period of time after the positive electricity voltage crest subsequently at significantly lower voltage and drops to negative voltage 818.HTG carries out modeling to biological neuron circuit principle, and this principle is that postsynaptic neuron is responsive to the excitant input at first, but the sensitivity of postsynaptic neuron rapidly disappears along with the excitant input that prolongs.Next a kind of possibility Mathematical Modeling of HTG is provided:
Z wherein
iIt is the gain that puts on input signal by the custom transmission gate;
T is the time;
y
iIt is the input of custom transmission gate;
T is a correction function;
F ([T (y
i)]) be the nonlinear function of calibrated input; And
A and B are constants.
In this equation, z
iIt is the gain that puts on the output signal of HTG unit with respect to input signal.In this model, the scope of gain is 0.0-1.0.First A (1-z
i) restoring relatively for a long time of HTG gain is provided, and second-Bf [T (y
i)] z
iWhen being provided at positive signal and inputing to HTG fast attenuated output signal and subsequently than the long duration negative signal.In representing any specific neuromorphic integrated circuit of one embodiment of the invention, the constant A of the specific mathematical model that adopts to any given HTG can be different with B, and perhaps the overall mathematical formulae of HTG can difference.Yet generally speaking, the HTG unit is used for weakening by nerve cell output or the prolongation excitant that receives or the effect of inhibition signal.
Fig. 9 is illustrated in the example that comprises the input computing unit and the computing unit of two kinds of other types of output computing unit that is adopted in hybrid microscale level-nanoscale neuromorphic integrated circuit of representing the embodiment of the invention.Input computing unit 902 is from being positioned at the signal transfer entity receiving inputted signal 904 outside hybrid microscale level-nanoscale neuromorphic integrated circuit, and handling this input signal after by the treated signal 906,908 of output signal pins 910,912 outputs.Output computing unit 914 is through outside computing unit reception excitant signal 916 and the inhibition signal 918 of input signal pin 920,922 from hybrid microscale level-nanoscale neuromorphic integrated circuit, and the processing input signal exports the treated signal 924 of the outside signal receiving entity of neuromorphic integrated circuit to generation.The output computing unit can receive input through semiconductor integrated circuit layer signal line separately, and the input computing unit can send signal to semiconductor integrated circuit layer signal line separately.In addition, input computing unit and output computing unit can be used to change between two kinds of different internal simulation signals in hybrid microscale level-nanoscale neuromorphic integrated circuit.
Figure 10 illustrates the input and output signal of the input computing unit of micron order-nanoscale composite nerve form circuit of representing one embodiment of the invention.Under many situations, input signal 1002 can be based on the digital signal of grouping, wherein comprises the regular length of sequence of binary digits or the grouping of variable-length and transfers to the input computing unit from external entity.The input computing unit is carried out the digital-to-analogue conversion of signals to produce aanalogvoltage or the current signal shown in signal relative time curve 1004 among Figure 10.Equally, output computing unit (914 among Fig. 9) receives analog signal and carries out the modulus signal conversion to export packet-based digital signal 1002.The digital signal that has many different possibility types, and the multiple different realizations of neuromorphic circuit can be adopted dissimilar analog signals.Therefore, the internal circuit of input computing unit and output computing unit quite depend on by hybrid microscale level-nanoscale neuromorphic integrated circuit receive or the type of the digital signal of output and comprehensive neuromorphic integrated circuit in polytype analog signal of being adopted.
Figure 11 A-B is illustrated in the interconnection of the computing unit in micron order-nanoscale neuromorphic integrated circuit of representing one embodiment of the invention.Figure 11 A illustrates 4 pin computing units of 3 * 3 arrays.As stated, each computing unit such as computing unit 1102 comprises two output pins 1104 and 1106, inhibition input pin 1108 and excitant input pin 1110.Figure 11 B illustrates the interconnection layer that is formed with two sub-layer that comprise parallel nanowires and memory resistor material sub-layers on the computing unit of 3 * 3 arrays shown in Figure 11 A.In Figure 11 B, such as each input pin of the input pin 1110 of computing unit 1102 with left-hand side be engaged to right-hand side near the nano wire 1114 of level link to each other through interface near the nano wire 1116 of level and with the pad 1112 that left-hand side and right-hand side nano wire 1114,1116 are engaged to input pin 1112.Therefore, be connected to first sublayer of whole nano wires formation parallel nanowires of the input pin in the computing unit array.Shown in Figure 11 B, nano wire rotates with respect to the direction of the upper and lower horizontal edge 1118,1120 of 3 * 3 array computation unit slightly.This rotation allows nano wire along direction horizontal-extending and to the right left, and strides across many neighborhood calculation unit and do not cover and be positioned at the inner or outside any additional vertical pins of computing unit, and nano wire is connected to computing unit through pad and vertical pins.Be connected to the nano wire of near vertical separately similarly such as the output pin of the output pin in the computing unit 1,102 1104.Therefore, the nano wire that is connected to the output pin of 3 * 3 array computation unit forms second sublayer that is close to parallel nanowires, and wherein the nano wire near vertical of second sublayer is in the nano wire of first sublayer.
In Figure 11 B, the memory resistor nanowire-junction between the nano wire is illustrated as the little filling dish of infall between two nano wires, such as filling dish 1124.The cynapse modeling of 1124 pairs of interconnection of memory resistor nanowire-junction presynaptic nerve cell 1126 and postsynaptic nerve cell 1128.Memory resistor nanowire-junction 1124 is with inhibition input pin 1132 interconnection of output pin 1130 with the postsynaptic nerve cell 1128 of presynaptic computing unit 1126.According to the present invention, a plurality of nanowire interconnections layers can be realized on the semiconductor integrated circuit layer of hybrid microscale level-nanoscale neuromorphic integrated circuit.A plurality of interconnection layers allow nerve cell on a plurality of layering logic aspects, to interconnect each other through cynapse class memory resistor nanowire-junction.Many interconnection layers neuromorphic integrated circuit framework of the present invention provides the quite a large amount of difference of computing unit maybe interconnection structure, and therefore provides very flexible with powerful interconnection structure with the very a large amount of different possible neuromorphic circuit of realization.
In some hybrid microscale level-nanoscale neuromorphic integrated circuit embodiment of the present invention; Can be in manufacture process configurable nanowire junctions or be ON and OFF state with its later programmed; Have only those nanowire-junction that are configured to ON through electric current and show the cynapse class behavior, and the nanowire-junction that is configured to OFF is served as open switch.In other hybrid microscale level-nanoscale neuromorphic integrated circuit embodiment of the present invention, nanowire-junction all is configured to the ON state, and the electricity of each nanowire-junction is led by the voltage signal through it and confirmed specially.
Figure 12 A-F illustrates the manufacturing of hybrid microscale level-nanoscale neuromorphic integrated circuit of representing one embodiment of the invention.Figure 12 A-F illustrates the manufacturing of neuromorphic integrated circuit to single computing unit.Yet as stated, hybrid microscale level-nanoscale neuromorphic integrated circuit comprises very a large amount of computing units, and they pave into the surface of hybrid microscale level-nanoscale neuromorphic integrated circuit together.In certain embodiments, can on every square centimeter of Semiconductor substrate, make several ten million and even several hundred million corresponding computing units.Manufacturing approach shown in Figure 12 A-F generally is applied to whole computing units of hybrid microscale level-nanoscale neuromorphic integrated circuit simultaneously.
Shown in Figure 12 A, each computing unit 1202 is to use any one manufacturing in the multiple method for manufacturing integrated circuit, comprises photoetching, doping, etching and planarization technique.The surface of computing unit 1204 comprises four conductive substrates 1206,1209 that aforesaid four vertical conduction pins are used.Computing unit can have any one in the multiple shape and size, and this depends on to the manufacturing approach and the Mathematical Modeling that are realized required computing unit behavior by application-specific.That the Common Shape of unit comprises is square, rectangle, hexagon and equilateral triangle, and they all can be scattered in two-dimensional array to pave into the surface of Semiconductor substrate fully.Shown in Figure 12 B; In the first step after making the bottom integrated circuit; First sublayer 1210 of nano wire is formed on the surface of integrated circuit by stamp or with other method; Every nano wire passes through for example vertical pins of shim-like structure contact of pad 1212, and stretches out to form a nano wire from pad along two rightabouts.Shown in Figure 12 B, the nano wire of first sublayer is parallel in all nano wires, and is connected to two input pins or two output pins of bottom computing unit, is connected to a pin at the most wherein for a nano wire of stator layers.The vertical pins that is not attached to the first sublayer nano wire is extended in first sublayer 1214 and 1216 as the path that covers bottom vertical pins conductive substrates (1207 and 1209 among Figure 12 A).
Then, shown in Figure 12 C, memory resistor material thin-layer 1220 is put on the upper surface of the first nano wire sublayer 1210.This memory resistor material forms the memory resistor nanowire-junction between the second sublayer nano wire that adds in the first sublayer nano wire 1210 and the subsequent step.Notice that conductive pin passes through the memory resistor material layer continuously through flush type path 1222-1225.Then, shown in Figure 12 D, second sublayer 1230 of nano wire is added to memory resistor layer 1220 top.The nano wire of nano wire second sublayer is parallel, and every nano wire is connected to single vertical pins.When the first nano wire sublayer was connected to output pin, the second nano wire sublayer was connected to input pin, and vice versa.Note the nano wire of second sublayer and the nano wire approximate vertical of first sublayer.Then, shown in Figure 12 E, add insulating barrier 1240 to the second nano wire sublayer 1230 tops, its vertical pins is extended through flush type path 1242-1245 equally.
Shown in Figure 12 F, additional nanowire interconnections layer 1250 can be formed on the top of first insulating barrier 1240.Represent micron order-nanoscale composite nerve form integrated circuit of all embodiment of the present invention can comprise the interconnection layer of any amount in fact, its top insulating barrier 1252 insulate the bottom interconnection layer basically fully and does not comprise that therefore usefulness is so that vertical pins extends to the flush type path of last surface of insulating layer.In addition, Figure 12 A-F illustrates manufacturing process from the angle of single computing unit, but concerning practical devices, in the series of steps shown in Figure 12 A-F on every square centimeter of substrate surface with millions of, several ten million or several hundred million computing units be manufactured on together.
Figure 13 illustrates the two-way signaling through the nano wire transmission of the nanowire interconnections layer of hybrid microscale level-nanoscale neuromorphic integrated circuit of representing one embodiment of the invention.In Figure 13, first nerves cell 1302 is via first nano wire 1306, second nano wire 1308 and with the memory resistor nanowire- junction 1310 and 1304 interconnection of nervus opticus cell of first nano wire 1306 and 1308 interconnection of second nano wire.First nano wire is connected to the output pin 1312 of the first presynaptic nerve cell 1302, and second nano wire 1310 is connected to the input pin 1314 of the second postsynaptic nerve cell 1304.As stated, when the presynaptic nerve cell had greater than active active of threshold value, the presynaptic nerve cell was emitted among Figure 13 the forward signal by direction arrow 1316 expressions.Equally, when the postsynaptic nerve cell has when being higher than the second threshold value activity level active, this postsynaptic nerve cell sends among Figure 13 and to pass through the reverse signal of input pin to presynaptic nerve cell 1302 by arrow 1318 expressions.Being transmitted in of the forward signal of nerve cell and reverse signal all provides in the certain embodiments of the invention as stated the state variation in the memory resistor material of the nanowire-junction of cynapse behavior modeling.
Figure 14 illustrates the electricity of the memory resistor nanowire-junction of presynaptic neuron and postsynaptic neuron interconnection is led variation.In Figure 14, each signal curve is with respect to the time vertical alignment, under all scenario with respect to for example the time during level of axle 1402 axle draw and form.Each curve curve among Figure 14 illustrates the signal and the state variation of the memory resistor nanowire-junction 1404 that first holding wire 1406 is connected with secondary signal line 1408.First curve 1410 illustrates the forward signal through 1406 transmission of first holding wire, and this forward signal is sent from the presynaptic node (1302 Figure 13) that is connected to first holding wire.Second curve 1412 illustrates the reverse signal through the transmission of secondary signal line, and this reverse signal sends from the postsynaptic node (1304 Figure 13) that is connected to the secondary signal line.The 3rd curve 1414 is illustrated in the voltage drop of each time point memory resistor 1404 both sides, and final curves 1416 are illustrated in the inductance of each time point memory resistor.For example the vertical dotted line of vertical dotted line 1418 illustrates each some time point in each curve 1410,1412,1414 and 1416.
When forward signal shows positive voltage crest 1422 or negative voltage crest 1424 and reverse signal when smooth, relatively little negative electricity pressure drop 1426 and relatively little positive voltage drops 1428 occur in the memory resistor both sides.These small voltage drop cause the memory resistor electricity to lead corresponding slightly minimizing 1430 and increase by a small margin 1432.When the voltage crest 1434,1436,1438,1440 of same-sign and equal magnitude all occurs in forward and the reverse signal, do not produce voltage drop 1442,1444 in memory resistor nanowire-junction both sides, so the electricity of memory resistor is led constant.Yet, when positive crest 1450 occur in the forward signal and negative peak 1452 same instantaneous when occurring in the reverse signal, big relatively negative electricity pressure drop 1454 occurs in the memory resistor both sides, this causes electricity to lead and correspondingly declines to a great extent 1456.Equally, in the forward signal time go up with reverse signal in positive crest 1462 corresponding negative voltage crests 1460 cause the big relatively positive voltage drops in memory resistor both sides 1464, this causes the corresponding significantly electricity of memory resistor material to lead increase by 1466.Therefore, through forward and the reverse signal that produces some type, nerve cell can make connection memory resistor nanowire-junction show like the electricity to the cynapse behavior modeling of mathematical expression in the preceding text and lead variation.Forward and reverse signal are used suitable waveform, can cause the memory resistor material to show the big variation that electricity is led when the postsynaptic, nerve had high activity after causing transmission following the high activity of forward signal closely through the postsynaptic nerve cell according to the Hebbian learning model of cynapse behavior.In addition, the conductivity of memory nanowire-junction reflects the passing pattern of activity level and launches from the signal of presynaptic that interconnects mutually and postsynaptic nerve cell.
All embodiment can make the multiple difference in functionality module that comprises a plurality of neighborhood calculation unit in hybrid microscale level-nanoscale neuromorphic integrated circuit according to the present invention.Figure 15 illustrates the realization of six computing unit dipoles according to an embodiment of the invention, and it shows the simulation behavior that is similar to digital logic flip-flops.Dipole module 1502 is fully made in the integrated circuit layer of hybrid microscale level-nanoscale neuromorphic integrated circuit of representing one embodiment of the invention.Nerve cell 1504,1506 is represented the input node of dipole, and nerve cell 1508,1510 is represented the output node of dipole.In dipole, an output node often has high activity when another output node has low activity, and vice versa, and the activated state of output node often keeps constant when the input node being had constant input or do not import.Therefore, dipole is as analog trigger, and locking lives in two kinds of output node a kind of in maybe the activity level states in fact.Input node 1504 is sent to output node 1508 via holding wire 1512 with the excitant signal, and via holding wire 1514 the inhibition signal is sent to output node 1510.Input node 1506 is followed similar but opposite signal transfer mode.Output node exports positive feedback signal 1518,1520 to input node 1504,1506 through HTG node 1524,1526.Therefore, when input node 1504 had high activity, output node 1408 often had high activity, and strengthened the high activity of input node 1504.When input node 1506 had high activity, output node 1510 often had high activity, and strengthened the high activity of input node 1506.Therefore left column computing node 1504 and 1508 fights for the high activity states with right column count node 1506,1510.
Figure 16 illustrates can be in many computing units module of second type of representing the hybrid microscale level of one embodiment of the invention-nanoscale neuromorphic integrated circuit to adopt.Module shown in Figure 16 be a kind of nine computing units around structure, wherein eight peripheral computing unit 1602-1609 send to central computation unit 1610 with inhibition branch line signal when work.Can in representing the hybrid microscale level-nanoscale neuromorphic integrated circuit of one embodiment of the invention, make comprise multiple varying number computing unit number of different types around structure.Under some situation, small-sized neighborhood calculation cell rings can be sent to central location with the excitant signal, and bigger peripheral computing unit ring can send to central computation unit with the inhibition signal, and vice versa.Other type around structure in, can feedback signal be sent to peripheral cell from central location, vice versa.Under some situation, a plurality of central locations can be surrounded by the one or more rings around the unit, and each ring provides the forward signal feedback of dissimilar or level.Be used to realize various types of mode identificating circuits and orientation maps continually around structure.
Figure 17 illustrates the layering interconnection of the computing unit in hybrid microscale level-nanoscale neuromorphic integrated circuit of representing one embodiment of the invention.Figure 17 illustrates 24 * 28 arrays of computing unit 1702.Logical layer is distributed to according to the logical layer key 1704 that the array below provides in each unit.For example, form first logical layer such as the shade computing unit that adds that adds shade computing unit 1706.The layering logic configuration of this computing unit can realize through the nerve cell that uses every layer of a nanowire interconnections layer interconnection.For example, the ground floor computing unit can be through nano wire in the first nanowire interconnections layer and the interconnection of memory resistor nanowire-junction side direction.The second logical layer unit can be similarly through the interconnection of the second nanowire interconnections layer.In addition, forward and feedback interconnection can be passed through a plurality of interconnection layers and therefore between all logical layers, handshaking be provided.The computing unit layer of layer sorting is applicable to polytype pattern recognition neuromorphic circuit and the inference machine of inferring from a plurality of inputs extractions.
Figure 18 illustrates the logical construction of second type that can in representing the hybrid microscale level-nanoscale neuromorphic integrated circuit of one embodiment of the invention, realize.Stratiform cortex shown in Figure 180 can comprise a plurality of lower floors transducers input computing unit such as computing unit 1802, and they for example all are connected to the dipole module to the dipole module 1804 of the individual unit modeling of stratiform cortex.The dipole module interconnect each other and with such as the interconnection of the complex cell of complex cell 1806, this complex cell exports signal to more high-rise computing unit or output unit.The stratiform cortex is found the purposes in the neuromorphic circuit of perceptron and many other types.Figure 19 illustrates stratiform cortex module, and this stratiform cortex module can be constructed in accordance and can be used to realize stratiform cortex neural form circuit with several adjacent stratiform cortex modules.Stratiform cortex module 1902 comprises six computing unit dipoles, and it is realized by six computing unit 1904-1909.The input of input computing unit 1910 representative sensors, and computing unit to 1912,1914 represent the stratiform cortex complex cell.Bigger module can comprise that the additional sensor input unit is to offer the dipole that comprises computing unit 1904-1909 with a plurality of transducer inputs.Interconnection between the unit can produce in the semiconductor integrated circuit layer specially, or can adopt one or more nanowire interconnections layers separately.
Although invention has been described to specific embodiment, yet the present invention is not intended to be subject to these embodiment.Modification in the spirit of the present invention is tangible to those skilled in that art.For example, computing unit also can adopt multiple cell type except that neuron, input unit, output unit and HTG unit.Although 4 pin computing units appear as many kinds of neuromorphic circuit enough connectivities are provided, yet in each alternate embodiment, possibly adopt the input and output pin of greater number or smaller amounts.As previously mentioned, can in hybrid microscale level-nanoscale neuromorphic integrated circuit, make additional signals line and other signal transmission associated component according to the embodiment of the invention.Can adopt semiconductor integrated circuit layer signal line that the computing unit of the many computing units module in hybrid microscale level-nanoscale neuromorphic integrated circuit and the extra connection between other logical block are provided.Semiconductor integrated circuit layer in hybrid microscale level-nanoscale neuromorphic integrated circuit can use in a large amount of ic manufacturing technologies any in any of number of different types Semiconductor substrate, to form.Neuromorphic integrated circuit can manufacture different sizes and area, and comprises additional function, comprises that defects detection, defective are improved and defective tolerance function is calculated the defective computing unit in the cell array with adjusting.Neuromorphic integrated circuit can adopt electric current and voltage signal in inside, and receives and export polytype external signal, comprises the signal based on digital packet.In many embodiment of the present invention, intrinsic nerve form integrated circuit signal comprises analog current and analog voltage signal.Although can in neuromorphic integrated circuit, adopt the polytype forward and the inverse analog signal mode of the biological aixs cylinder spike of simulation, yet neuromorphic integrated circuit also can be based on simple voltage level and/or the work of current level signal.Neuromorphic integrated circuit can comprise the for example additional function of clock signal transfer function, timing signal is offered input computing unit and output computing unit.Neuromorphic integrated circuit can comprise a plurality of integrated circuit layers separately, each integrated circuit layer and related a plurality of nanowire interconnections layer interconnection.
For the ease of explaining, the explanation of front uses particular term to provide thorough of the present invention.Yet, it is apparent that concerning those skilled in that art it is essential to the invention that these details are not realization.The front is described as purpose to the explanation of specific embodiment of the present invention to separate to mediate a settlement.They be not exhaustive or the present invention is defined in form accurately.Many modifications and variation are feasible in view of above-mentioned religious doctrine.The description of embodiment and explanation are in order to explain the principle that use the present invention and its practical field with best way, make those skilled in that art can utilize the present invention with each embodiment and visualize the various correction forms that are suitable for special purpose best thus.Scope of the present invention is limited following claims and equivalent thereof.
Claims (14)
1. neuromorphic integrated circuit comprises:
The analogue computer array of on the integrated circuit substrate, making, the Analogical Electronics in each computing unit are connected to one or more first kind pins and one or more second type of pin, and these pins generally perpendicularly stretch out from said computing unit; And
One or more nanowire interconnections layers above said analogue computer array, each nanowire interconnections layer comprises:
First sublayer of almost parallel nano wire, every nano wire of said first sublayer are connected to the single computing unit pin of the first kind;
The memory resistor sublayer;
Second sublayer of almost parallel nano wire, every nano wire of said second sublayer are connected to the single computing unit pin of second type, and the nano wire of said second sublayer is along the direction orientation of the nano wire that is not parallel to said first sublayer; And
The nano wire of the nano wire of said second sublayer and said first sublayer overlaps, thereby forms the memory resistor nanowire-junction, each nanowire-junction through to the said memory resistor sublayer of cynapse modeling with the second sublayer nano wire and the first sublayer nanowire interconnections.
2. neuromorphic integrated circuit as claimed in claim 1 is characterized in that the memory resistor nanowire-junction realizes a kind of model, and said model can be expressed as:
Z wherein
IjBe weight or the gain that is produced through cynapse ij with presynaptic neuron i and postsynaptic neuron j interconnection;
ε is a pace of learning;
ω forgets speed;
x
iIt is the activity of neuron i;
x
jIt is the activity of neuron j;
F ((x
j) (ω z
Ij+ g (x
i))) be that variable is (x
j) (ω z
Ij+ g (x
i)) nonlinear function;
G (x
i) be the active nonlinear function of neuron i; And
T is the time.
3. neuromorphic integrated circuit as claimed in claim 1 is characterized in that, said computing unit comprises:
The neuron computes unit;
The input computing unit; And
The output computing unit.
4. neuromorphic integrated circuit as claimed in claim 3 is characterized in that, each neuron computes unit comprises:
With the first summation function assembly input of excitant signal and that export said neuron computes unit to;
With the second summation function assembly input of inhibition signal and that export said neuron computes unit to;
To merge signal generation functional unit from the output of the said first and second summation function assemblies with the output analog signal;
In the last time interval, the output of said signal generation functional unit is quadratured with the leaky functional unit of signal behind the output integration; And
When signal exceeds a threshold value behind the said integration, activate and get the threshold function assembly from the output signal emission of said neuron computes unit.
5. neuromorphic integrated circuit as claimed in claim 3 is characterized in that, a kind of model is realized in said neuron computes unit, and said model can be expressed as:
Wherein i is the neuron through modeling;
x
iIt is the activity of neuron i;
T is the time;
f
j(x
j) be the nonlinear function of the activity of neuron j;
z
JiIt is the weight of the cynapse between neuron j and the i;
g
j(x
j) be the nonlinear function of the activity of neuron j;
N is the neuron number that links through excitant input and node i;
M is the neuron number that links through inhibition input and node i;
A and B are constants.
6. neuromorphic integrated circuit as claimed in claim 3 is characterized in that, each input computing unit comprises:
Receive external signal and the conversion of signals that is received become to output to the coding circuit of analog signal of one or more computing units of said neuromorphic integrated circuit.
7. neuromorphic integrated circuit as claimed in claim 6 is characterized in that said external signal is based on the signal of digital packet, and said coding circuit becomes one of analog voltage signal or analog current signal with said signal transition.
8. neuromorphic integrated circuit as claimed in claim 3 is characterized in that, each output computing unit comprises:
Decoding circuit; The analog-signal transitions that said decoding circuit will input to the analogue signal generating of said output computing unit from the one or more neuron computes unit through said neuromorphic integrated circuit becomes external signal, and said external signal is exported to the outside entity of said neuromorphic integrated circuit.
9. neuromorphic integrated circuit as claimed in claim 8 is characterized in that said external signal is based on the signal of digital packet.
10. neuromorphic integrated circuit as claimed in claim 1 is characterized in that, said computing unit also comprises custom transmission gate unit.
11. neuromorphic integrated circuit as claimed in claim 10; It is characterized in that; Two computing units of holding wire interconnection in each integrated circuit layer of being accustomed to using said neuromorphic integrated circuit in transmission gate unit, thus the gain of will doubling puts on input signal to produce the output signal.
12. neuromorphic integrated circuit as claimed in claim 10 is characterized in that, a kind of model is realized in each custom transmission gate unit, and said model can be expressed as:
Z wherein
iIt is the gain that puts on input signal by the custom transmission gate;
T is the time;
y
iIt is the input of custom transmission gate;
T is a correction function;
F ([T (y
i)]) be the nonlinear function of calibrated input; And
A and B are constants.
13. neuromorphic integrated circuit as claimed in claim 10 is characterized in that, two or more computing units combine to form module, comprising:
Dipole;
Around structure; And
Stratiform cortex subelement.
14. neuromorphic integrated circuit as claimed in claim 10 is characterized in that, said computing unit is organized into a plurality of hierarchy layers, and the unit of each hierarchy layer is through intercom with the corresponding nanowire interconnections layer of said hierarchy layer mutually.
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US99266307P | 2007-12-05 | 2007-12-05 | |
| US60/992,663 | 2007-12-05 | ||
| PCT/US2008/006592 WO2009075694A1 (en) | 2007-12-05 | 2008-05-22 | Hybrid microscale-nanoscale neuromorphic integrated circuit |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101889343A CN101889343A (en) | 2010-11-17 |
| CN101889343B true CN101889343B (en) | 2012-11-21 |
Family
ID=40755768
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2008801196638A Expired - Fee Related CN101889343B (en) | 2007-12-05 | 2008-05-22 | Hybrid microscale-nanoscale neuromorphic integrated circuit |
Country Status (6)
| Country | Link |
|---|---|
| US (1) | US8332340B2 (en) |
| JP (1) | JP2011507232A (en) |
| KR (1) | KR101438469B1 (en) |
| CN (1) | CN101889343B (en) |
| DE (1) | DE112008003510T5 (en) |
| WO (1) | WO2009075694A1 (en) |
Families Citing this family (52)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9679242B2 (en) | 2002-03-12 | 2017-06-13 | Knowm Tech, Llc | Memristor apparatus with meta-stable switching elements |
| US9104975B2 (en) * | 2002-03-12 | 2015-08-11 | Knowmtech, Llc | Memristor apparatus |
| KR101002683B1 (en) * | 2008-08-19 | 2010-12-20 | 한국기계연구원 | High pressure ball valve within the double seat ring and relief valve |
| CN102803981B (en) | 2009-06-19 | 2016-09-21 | 皇家飞利浦电子股份有限公司 | Memristor device is used for MRI RF coil |
| US8575585B2 (en) * | 2009-07-13 | 2013-11-05 | Hewlett-Packard Development Company, L.P. | Memristive device |
| US8275728B2 (en) * | 2009-11-05 | 2012-09-25 | The United States Of America As Represented By The Secretary Of The Air Force | Neuromorphic computer |
| US8311965B2 (en) * | 2009-11-18 | 2012-11-13 | International Business Machines Corporation | Area efficient neuromorphic circuits using field effect transistors (FET) and variable resistance material |
| US8716688B2 (en) * | 2010-02-25 | 2014-05-06 | The University Of Kentucky Research Foundation | Electronic device incorporating memristor made from metallic nanowire |
| WO2011155928A1 (en) * | 2010-06-08 | 2011-12-15 | Hewlett-Packard Development Company, L.P. | Three-dimensional integrated circuit |
| US8433665B2 (en) * | 2010-07-07 | 2013-04-30 | Qualcomm Incorporated | Methods and systems for three-memristor synapse with STDP and dopamine signaling |
| US9092736B2 (en) * | 2010-07-07 | 2015-07-28 | Qualcomm Incorporated | Communication and synapse training method and hardware for biologically inspired networks |
| US9293200B2 (en) | 2010-08-30 | 2016-03-22 | Hewlett Packard Enterprise Development Lp | Multilayer memory array |
| US8510239B2 (en) * | 2010-10-29 | 2013-08-13 | International Business Machines Corporation | Compact cognitive synaptic computing circuits with crossbar arrays spatially in a staggered pattern |
| US8655813B2 (en) * | 2010-12-30 | 2014-02-18 | International Business Machines Corporation | Synaptic weight normalized spiking neuronal networks |
| US8848337B2 (en) | 2011-02-01 | 2014-09-30 | John R. Koza | Signal processing devices having one or more memristors |
| KR101888468B1 (en) | 2011-06-08 | 2018-08-16 | 삼성전자주식회사 | Synapse for a function cell of spike-timing-dependent plasticity(stdp), the function cell of spike-timing-dependent plasticity, and a neuromorphic circuit using the function cell of spike-timing-dependent plasticity |
| FR2978271B1 (en) * | 2011-07-21 | 2014-03-14 | Commissariat Energie Atomique | DEVICE AND METHOD FOR PROCESSING DATA |
| US9262712B2 (en) * | 2013-03-08 | 2016-02-16 | International Business Machines Corporation | Structural descriptions for neurosynaptic networks |
| KR102179899B1 (en) | 2013-08-05 | 2020-11-18 | 삼성전자주식회사 | Neuromophic system and configuration method thereof |
| US20150220831A1 (en) * | 2014-02-06 | 2015-08-06 | Qualcomm Incorporated | Short-term synaptic memory based on a presynaptic spike |
| US9195903B2 (en) | 2014-04-29 | 2015-11-24 | International Business Machines Corporation | Extracting salient features from video using a neurosynaptic system |
| US9373058B2 (en) | 2014-05-29 | 2016-06-21 | International Business Machines Corporation | Scene understanding using a neurosynaptic system |
| US9978015B2 (en) * | 2014-05-30 | 2018-05-22 | Knowmtech, Llc | Cortical processing with thermodynamic RAM |
| US9798972B2 (en) | 2014-07-02 | 2017-10-24 | International Business Machines Corporation | Feature extraction using a neurosynaptic system for object classification |
| US10115054B2 (en) | 2014-07-02 | 2018-10-30 | International Business Machines Corporation | Classifying features using a neurosynaptic system |
| US10489700B1 (en) * | 2014-07-07 | 2019-11-26 | Crossbar, Inc. | Neuromorphic logic for an array of high on/off ratio non-volatile memory cells |
| KR102255309B1 (en) * | 2014-10-17 | 2021-05-25 | 삼성전자주식회사 | Artificial neuron semiconductor device with three dimension structure and artificial neuron semiconductor system using the same |
| US9852370B2 (en) | 2014-10-30 | 2017-12-26 | International Business Machines Corporation | Mapping graphs onto core-based neuromorphic architectures |
| US9704094B2 (en) | 2015-02-19 | 2017-07-11 | International Business Machines Corporation | Mapping of algorithms to neurosynaptic hardware |
| US10204301B2 (en) | 2015-03-18 | 2019-02-12 | International Business Machines Corporation | Implementing a neural network algorithm on a neurosynaptic substrate based on criteria related to the neurosynaptic substrate |
| US9971965B2 (en) | 2015-03-18 | 2018-05-15 | International Business Machines Corporation | Implementing a neural network algorithm on a neurosynaptic substrate based on metadata associated with the neural network algorithm |
| CN104701309B (en) * | 2015-03-24 | 2017-10-13 | 上海新储集成电路有限公司 | Three-dimensional stacked formula neuron device and preparation method |
| US9984323B2 (en) | 2015-03-26 | 2018-05-29 | International Business Machines Corporation | Compositional prototypes for scalable neurosynaptic networks |
| CN105740946B (en) * | 2015-07-29 | 2019-02-12 | 上海磁宇信息科技有限公司 | A kind of method that application cell array computation system realizes neural computing |
| CN105404925A (en) * | 2015-11-02 | 2016-03-16 | 上海新储集成电路有限公司 | Three-dimensional nerve network chip |
| WO2017105396A1 (en) * | 2015-12-14 | 2017-06-22 | Intel Corporation | Magnetoelectric oscillatory cellular neural network apparatus and method |
| US10650309B2 (en) * | 2016-08-09 | 2020-05-12 | International Machines Corporation | High dynamic range, high class count, high input rate winner-take-all on neuromorphic hardware |
| FR3059805B1 (en) * | 2016-12-02 | 2019-03-15 | Commissariat A L'energie Atomique Et Aux Energies Alternatives | ARTIFICIAL SYNAPSE MODULATION DEVICE AND METHOD COMPRISING SAID MODULATION DEVICE, SHORT-TERM PLASTICITY METHOD IN ARTIFICIAL NEURON NETWORK COMPRISING SAID ARTIFICIAL SYNAPSE |
| US10235565B2 (en) * | 2016-12-21 | 2019-03-19 | Volkswagen Ag | System and methodologies for occupant monitoring utilizing digital neuromorphic (NM) data and fovea tracking |
| US10229341B2 (en) | 2016-12-21 | 2019-03-12 | Volkswagen Ag | Vector engine and methodologies using digital neuromorphic (NM) data |
| US10133944B2 (en) | 2016-12-21 | 2018-11-20 | Volkswagen Ag | Digital neuromorphic (NM) sensor array, detector, engine and methodologies |
| US10282615B2 (en) | 2016-12-21 | 2019-05-07 | Volkswagen Ag | System and method for root association in image data |
| US10789495B2 (en) | 2016-12-21 | 2020-09-29 | Volkswagen Ag | System and method for 1D root association providing sparsity guarantee in image data |
| KR102143440B1 (en) * | 2017-01-20 | 2020-08-11 | 한양대학교 산학협력단 | 3d neuromorphic device and method of manufacturing the same |
| US11308387B2 (en) * | 2017-05-09 | 2022-04-19 | Samsung Electronics Co., Ltd. | STDP with synaptic fatigue for learning of spike-time-coded patterns in the presence of parallel rate-coding |
| US11748607B2 (en) * | 2017-07-31 | 2023-09-05 | Syntiant | Systems and methods for partial digital retraining |
| KR102589968B1 (en) * | 2018-04-17 | 2023-10-16 | 삼성전자주식회사 | Neuromorphic circuit having 3D stacked structure and Semiconductor device having the same |
| US10748608B2 (en) | 2018-10-12 | 2020-08-18 | At&T Intellectual Property I, L.P. | Memristive device and method based on ion migration over one or more nanowires |
| US10922824B1 (en) | 2019-08-13 | 2021-02-16 | Volkswagen Ag | Object tracking using contour filters and scalers |
| KR20210050243A (en) | 2019-10-28 | 2021-05-07 | 삼성전자주식회사 | Neuromorphic package devices and neuromorphic computing systems |
| TWI728556B (en) * | 2019-11-18 | 2021-05-21 | 財團法人工業技術研究院 | Neuron circuit and artificial neural network chip |
| US11856877B2 (en) | 2019-12-23 | 2023-12-26 | The University Of Canterbury | Electrical contacts for nanoparticle networks |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5376813A (en) * | 1993-06-02 | 1994-12-27 | California Institute Of Technology | Adaptive photoreceptor including adaptive element for long-time-constant continuous adaptation with low offset and insensitivity to light |
| US5781648A (en) * | 1995-04-07 | 1998-07-14 | California Institute Of Technology | Pulse domain neuromorphic integrated circuit for computing motion |
| CN1447457A (en) * | 2002-03-27 | 2003-10-08 | 夏普株式会社 | Integrated circuit device and neure |
-
2008
- 2008-05-22 JP JP2010536903A patent/JP2011507232A/en active Pending
- 2008-05-22 WO PCT/US2008/006592 patent/WO2009075694A1/en active Application Filing
- 2008-05-22 DE DE112008003510T patent/DE112008003510T5/en not_active Ceased
- 2008-05-22 CN CN2008801196638A patent/CN101889343B/en not_active Expired - Fee Related
- 2008-05-22 KR KR1020107014868A patent/KR101438469B1/en active IP Right Grant
- 2008-05-22 US US12/743,781 patent/US8332340B2/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5376813A (en) * | 1993-06-02 | 1994-12-27 | California Institute Of Technology | Adaptive photoreceptor including adaptive element for long-time-constant continuous adaptation with low offset and insensitivity to light |
| US5781648A (en) * | 1995-04-07 | 1998-07-14 | California Institute Of Technology | Pulse domain neuromorphic integrated circuit for computing motion |
| CN1447457A (en) * | 2002-03-27 | 2003-10-08 | 夏普株式会社 | Integrated circuit device and neure |
Also Published As
| Publication number | Publication date |
|---|---|
| JP2011507232A (en) | 2011-03-03 |
| KR101438469B1 (en) | 2014-09-05 |
| KR20100099256A (en) | 2010-09-10 |
| WO2009075694A1 (en) | 2009-06-18 |
| US20100277232A1 (en) | 2010-11-04 |
| DE112008003510T5 (en) | 2010-11-04 |
| CN101889343A (en) | 2010-11-17 |
| US8332340B2 (en) | 2012-12-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN101889343B (en) | Hybrid microscale-nanoscale neuromorphic integrated circuit | |
| Basu et al. | Low-power, adaptive neuromorphic systems: Recent progress and future directions | |
| Krestinskaya et al. | Neuromemristive circuits for edge computing: A review | |
| Snider | Spike-timing-dependent learning in memristive nanodevices | |
| Snider | Self-organized computation with unreliable, memristive nanodevices | |
| CN1098503C (en) | Neural net with spatially distributed functionalities | |
| Yakopcic et al. | Memristor crossbar deep network implementation based on a convolutional neural network | |
| Srinivasan et al. | Restocnet: Residual stochastic binary convolutional spiking neural network for memory-efficient neuromorphic computing | |
| Jang et al. | Optimization of conductance change in Pr 1–x Ca x MnO 3-based synaptic devices for neuromorphic systems | |
| CN101971166B (en) | Neuromorphic circuit | |
| Velasquez et al. | Automated synthesis of crossbars for nanoscale computing using formal methods | |
| Gao et al. | Cortical models onto CMOL and CMOS—architectures and performance/price | |
| Strukov et al. | Memory technologies for neural networks | |
| Chen et al. | Universal perceptron and DNA-like learning algorithm for binary neural networks: LSBF and PBF implementations | |
| Bennett et al. | Supervised learning with organic memristor devices and prospects for neural crossbar arrays | |
| Xu et al. | Optimized near-zero quantization method for flexible memristor based neural network | |
| Schmid | Neuromorphic microelectronics from devices to hardware systems and applications | |
| Lee et al. | CMOL CrossNets as pattern classifiers | |
| Moposita et al. | Efficiency of double-barrier magnetic tunnel junction-based digital eNVM array for neuro-inspired computing | |
| Payvand et al. | Exploiting local connectivity of CMOL architecture for highly parallel orientation selective neuromorphic chips | |
| dos Santos Pês et al. | A Spiking Neural Network implemented with Single-Electron Transistors and NoCs | |
| Parmar et al. | Short-term plasticity circuit device exploration in the MASTISK neuromorphic framework | |
| Fu et al. | Cycle-to-cycle variation enabled energy efficient privacy preserving technology in ann | |
| Cha et al. | A geometrical optimization rule of the synaptic pass-transistor for a low power analog accelerator | |
| Voyles et al. | Soft robotic skin from intelligent meta-materials |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| C41 | Transfer of patent application or patent right or utility model | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20170110 Address after: Texas, USA Patentee after: HEWLETT PACKARD ENTERPRISE DEVELOPMENT LP Address before: Texas, USA Patentee before: Hewlett-Packard Development Company, L.P. |
|
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20121121 Termination date: 20210522 |